Self-Routing Capsule Networks

Taeyoung Hahn, Myeongjang Pyeon, Gunhee Kim

Abstract: Capsule networks have recently gained a great deal of interest as a new architecture of neural networks that can be more robust to input perturbation than similar-sized CNNs. Capsule networks have two distinctions from the conventional CNNs: (i) each layer consists of a set of capsules that specialize in disjoint regions of the feature space and (ii) the routing-by-agreement coordinates connections from the capsules' predictions in the previous layer to the capsules' inputs in the next layer. Although the routing-by-agreement is capable of filtering out the noisy predictions of capsules by dynamically adjusting their gating values, it also causes two side effects: (i) high computational complexity and (ii) training instability due to the unsupervised clustering. In this work, we design a novel routing strategy called self-routing where each capsule is routed independently by its subordinate routing network and thus the agreement between capsules is not required anymore. As a consequence, both poses and activations of upper-level capsules are obtained in a way similar to mixture-of-experts. Our experiments on CIFAR-10, SVHN, and SmallNORB show that the self-routing performs not only better for image classification but also more robust against white-box adversarial attacks and novel viewpoints by affine transformation while requiring less computation, compared to the previous agreement-based methods.

Code Link: https://github.com/coder3000/SR-CapsNet

CMT Num: 4168